2012 年 78 巻 789 号 p. 1886-1898
This paper studies locomotion patterns of biological robots by using Q-learning method. Previous studies showed that primitive locomotion patterns were obtained using basic motion rhythm given by its designer. On the other hand, this paper shows that the biological robots acquire its locomotion pattern by itself using a reinforcement learning method. Therefore, advantage of this study is that the biological robots can obtain optimum locomotion pattern related to both its mechanism and actuator characteristics, which means that this method can evaluate the biological robot's mechanism on the view point of the locomotion pattern. We evaluate three different mechanisms of the biological robots which are four legs with a vertebral column, four legs without the vertebral and 6 legs on the viewpoint of energetic efficiencies. The rewards are both advance distance and kinetic energy. Simulation results show interesting results similar to actual living things.